Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

2.5 Cutting the Fat

One of the fastest ways to waste tokens is through unnecessary detail or weak phrasing. A bloated Personality or Scenario not only eats into your budget but also dilutes the signals that actually drive behavior. The trick is to trim everything that doesn’t change how the bot acts.

Three Common Mistakes

Trivia Overload
Bad: “Her uncle worked in a steel mill in 1983, and her cousin once saw a UFO.”
Good: “She comes from a working-class family and carries their practical outlook.”
Why: The first example dumps trivia that will never surface in dialogue. The second compresses the idea into a behavioral cue the bot can express.

Biography Instead of Behavior
Bad: “Born in London in 1995, she attended St. Andrews…”
Good: “Her voice carries traces of a London upbringing, clipped vowels in stress.”
Why: A birthdate and school are census records. They don’t shape speech. The rewritten version shows how upbringing influences dialogue.

Hedging Words
Bad: “She might sometimes act shy if the situation feels right.”
Good: “She acts shy when meeting new people.”
Why: The words “might” and “sometimes” are read by the model as optional. Definitive phrasing turns the trait into a consistent behavior rule.

Passive phrasing is vague.
“Defined by shyness” → too abstract.

Active phrasing is concrete.
“Speaks softly and avoids eye contact” → gives the model tokens it can directly use to generate dialogue and body language.

Other Token Traps

  • Redundancy: Saying the same trait multiple ways (“She is quiet. She is reserved. She doesn’t talk much.”) can be collapsed into one stronger line.

  • Excess adjectives: “She is very, very shy” adds nothing. One “shy” is enough.

  • Soft hedges: “Sometimes, maybe, tends to…” weaken rules. Delete them.

Design Lesson

Every sentence in your prompt should either define voice, behavior, or relationship. If it doesn’t, it’s fat. Trim aggressively until only functional rules remain. The more concise the instructions, the sharper and more stable the bot’s performance will be.